EFS-IDS: An Enhanced Feature-Selective Intrusion Detection System for Imbalanced IoT Traffic Data

Abstract

The rapid proliferation of the Internet of Things (IoT) has expanded the surface of cyberattacks, intensifying the need for intelligent and adaptive Intrusion Detection Systems (IDS). Conventional IDS solutions often struggle with high-dimensional feature spaces, class imbalance, and computational limitations of edge and fog devices. This paper presents EFS-IDS, an Efficient Feature-Selection-based Intrusion Detection System tailored for IoT environments. The framework addresses three significant challenges: feature redundancy, data imbalance, and model generalization through a multistage deep learning pipeline. First, an ensemble feature selection mechanism combines Information Gain, Fast Correlation-Based Filter, and Average Feature Importance to identify a compact yet highly discriminative subset of features. Second, a hybrid balancing strategy integrating SMOTE and Borderline-SMOTE mitigates skewed class distributions, improving recognition of minority attacks. Third, a cost-sensitive CNN–DNN hybrid classifier leverages convolutional layers for localized flow pattern extraction and deep dense layers for global decision modeling, optimized through a weighted cross-entropy loss. Together, these modules enhance detection accuracy, robustness, and resource efficiency across heterogeneous IoT devices. Extensive experiments on the CIC-IDS and CIC-IoT benchmark datasets show that EFS-IDS achieves up to 98% accuracy and 0.96 F1-score, outperforming state-of-the-art models in both balanced and imbalanced conditions. The proposed framework demonstrates superior adaptability, reduced false alarms, and efficient deployment across edge–fog–cloud layers, positioning EFS-IDS as a scalable and effective defense mechanism for next-generation IoT networks.